Chinese AI labs are forcing the market to compete on the cost of useful intelligence. Export controls restricted Chinese labs’ access to advanced NVIDIA chips. Staying competitive meant extracting more performance from the hardware available. In training, Deepseek kept GPUs working while data moved between them and shifted to lower-precision formats that halve memory and compute. The bigger savings came in serving: sparse models activate only a fraction of their parameters per token, and redesigned attention cut the memory cost of long contexts. The result is models that are cheaper to build and far cheaper to run. Chinese open-weight models went from under 1.2% of weekly token consumption on OpenRouter in late 2024 to a majority in 2026. OpenRouter is a small percentage of global token volumes and its users are price-elastic so this should not be read as global market share. It does show that users will choose cheaper models when switching is easy and performance is good enough. Cheaper inference also has strategic value for China. These models can already run on domestic accelerators and reduce the country’s dependence on foreign hardware for inference. China still relies largely on foreign chips for frontier training. China will not have full compute independence until its labs can train frontier models at scale on domestic infrastructure.